Research On Unsupervised Person Re-identification Based On Deep Tracklet Association Learning

With the rapid development of society and economy,people pay more and more attention to urban public safety issues,therefore,video surveillance networks are widely deployed in every corner of the city.As an important part of intelligent video surveillance system,person re-identification has attracted more and more attention of the researchers.Person re-identification mainly refers to given a person image under a surveillance camera,using computer vision technology to determine whet her the person appears in the images captured by another surveillance camera with a different angle or field of view.Due to the rich spatiotemporal information of person video tracklets,many video-based person re-identification algorithms have been deve loped in recent years.However,existing methods are usually based on strongly supervised learning,which requires a large amount of person image identity information.Unsupervised learning does not require person image identity information,but the images of the same person are very different,and the difference between different person images is very small,which brings great difficulties to the research of unsupervised person re-identification.In order to solve this problem,this paper designs an unsupervised video-based person re-identification algorithm.The main research work and research results are as follows:Firstly,this paper proposes an unsupervised person re-identification algorithm based on tracklet association learning.This paper utilizes the sparse space-time tracklet sampling method to sample the person video tracklets in each camera in the video dataset.After sampling,each person has only one video tracklet under each camera.At the same time,this paper proposes a video feature memory module to store the video-level features of all the person video tracklets,and then performs intra-camera tracklet association learning and cross-camera tracklet association learning in turn.In the intra-camera tracklet association learning,this paper utilizes the affiliation relation between the person image and the person video tracklets to learn more discriminative information between persons in each camera.Then the exponential moving average method is utilized to gradually update the person image feature to the person video-level features in the video feature memory module.In the cross-camera tracklet association learning,this paper proposes a cyclic ranking alignment and similarity threshold filtering method to mine potential positive pairs of person video tracklets across cameras.Therefore,the cross-camera vision-invariant features are learned by pulling the potential positive pairs closer.Secondly,in order to extract more fine-grained attribute features of the persons,this paper uses the Transformer network architecture as the feature extraction network to extract the features of the person images.In this paper,the person image is divided into several blocks of the same size and the corresponding position embedding information and a learnable feature vector are added.The self-attention mechanism in the Transformer network architecture is used to make the extracted features contain both global attributes and local attributes.The local saliency features of the persons are mined to improve the expression ability of the extracted features.At the same time,more discriminative attribute features between sim ilar but different persons are obtained to complete the person re-identification task better.Finally,the experimental results on two large public video datasets,i.e.,MARS and Duke MTMC-Video Re ID datasets,show that the recognition results of this algorithm are better than other unsupervised person re-identification methods.The proposed method can effectively complete the person re-identification task,which is helpful for the development of the person re-identification.